A Comparative Analysis of Feature Selection Methods for Machine Learning-Based Phishing URL Detection

Abstract

This study investigates the impact of three categories of feature selection methods which are filter-based, wrapper-based, and embedded on the performance and efficiency of machine learning classifiers for phishing URL detection. Experiments were conducted on the PhiUSIIL Phishing URL Dataset comprising 235,795 URLs (134,850 legitimate, 100,945 phishing) with 20 URL structural features. Results demonstrate that LASSO (Embedded) at k = 5 achieves an F1 score of 0.9975 versus the 20-feature baseline of 0.9965 (ΔF1 = 0.001). Given that this difference falls within the per-fold standard deviation of both configurations (≤ 0.0003), we treat the two configurations as operationally equivalent and frame the contribution as efficiency at unchanged accuracy rather than as an accuracy improvement.